Abstract: The specification of hyperparameters plays a critical role in determining the practical performance of a machine learning method. Hyperparameter Optimization (HPO), i.e., the searching for optimal specification of hyperparameters, however, often faces critical computational challenges due to the vast searching space and the high computational cost on model training under a given hyperparameter specification. In this paper, we propose BOPT-HPO, a systematic approach for efficient HPO by leveraging Bayesian optimization with Pareto-principled training, based on the observation that the training procedure of a machine learning method under a given hyperparameter specification often follows the Pareto principle (the 80/20 rule) that about 80% of the total improvement in the objective function is achieved in 20% of the training time. By introducing two levels of training corresponding to the Pareto principle, i.e., the eighty-percent training (ET) and the complete training (CT), and establishing a joint surrogate model for CT runs and ET runs, BOPT-HPO reduces the computational cost of HPO significantly under the framework of Bayesian optimization with multi-fidelity measurements. A wide range of experimental studies confirm that the proposed approach achieves economical HPO for various machine learning models, including support vector machines, feed-forward neural networks, and convolutional neural networks.

Key words and phrases: Automated artificial intelligence, black-box function optimization, computer experiments, multi-fidelity modelling, truncated Gaussian process.